Research On Target-like Interference Suppression Algorithm Of Gesture Target By FMCW Radar

With the continuous rise of the fifth-generation(5G)and artificial intelligence,the development of various human-computer interaction(HCI)technologies,more and more researchers have begun to engage in the fields of artificial intelligence,autonomous driving and human-computer interaction.Gesture has gradually become a focus of current research as the most common communication method in HCI.Traditional gesture recognition methods mainly use wearable devices or camera devices for recognition.These methods can not be widely used because it is inconvenient to wear and easily affected by light intensity.Compared with traditional gesture recognition methods,gesture recognition based on Frequency Modulated Continuous Wave(FMCW)radar has the advantages of convenience and strong anti-interference.In this paper,we proposed the research of FMCW gesture recognition method based on target-like interference suppression.Firstly,we modify the phase offset of the intermediate frequency(IF)signal of the radar,and use the 2D Fast Fourier Transform(2D-FFT)algorithm and Multiple Signal classification(MUSIC)algorithm to extracts the range,velocity and angle parameters of the gesture target.We finally get a multi-frame Range-Doppler Map(RDM)and a single-frame angle time map(ATM).Secondly,we studied target-like interference suppression methods based on RDM and ATM.According to the magnitude of the target power after parameter estimation,we used the Constant False Alarm Rate(CFAR)to design an adaptive threshold,and used it to extract the target signal.Then we combined the target parameter information to correct the error between the radial velocity of the signal and the actual movement speed of the gesture,and improved Kalman filtering to keep the residual sequence orthogonal,predicted the trajectory of gesture targets and suppressed non-gesture targets.Finally,we studied multi-parameter gesture recognition methods.We constructed the range-doppler map and angle-time map.According to the characteristics of each data set,we used 3D Inflated Convolution Neural Network(I3D)and single-parameter neural network for feature extraction,and then each parameter are used to extract the temporal correlation features of gestures.Finally,the RDM features are segmented and recombined with ATM features.And Support Vector Machine(SVM)and Fully Connected Layers(FC)are used for feature fusion and classification.